LoRA for Gender-Inclusive Rewriting and Activation Steering for Counter-Narrative Generation

Akhil Rajeev P; Manoj Balaji J

REVIEW 3 major objections 6 minor 18 references

LoRA for Gender-Inclusive Rewriting and Activation Steering for Counter-Narrative Generation

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a single PCA-derived steering direction injected into one layer of Gemma-3-4B-it can shift counter-narrative outputs toward inclusive, polite language without weight updates, achieving an official score of 78.12%.

desk verdict Honest shared-task write-up, but the headline claim about activation steering is unsubstantiated—the missing no-steering baseline is the whole ballgame. read the letter →

arxiv 2607.23083 v1 pith:QJMV3V7K submitted 2026-07-25 cs.CL

Akhil Rajeev P , Manoj Balaji J This is my paper

classification cs.CL

keywords gender-inclusivelanguagecounter-narrativegenerationactivationsteeringrepresentationengineeringPCALoRAfine-tuninginference-timeinterventionGemma-3

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that inference-time activation steering can serve as a compute-efficient alternative to fine-tuning for socially aligned language generation. The authors derive a principal steering direction from contrastive hidden-state activations and inject it into a single mid-layer of Gemma-3-4B-it, reporting competitive scores for gender-inclusive counter-narrative generation. If correct, this would let practitioners redirect model behavior without the cost of retraining. The paper also catalogues concrete failure modes, showing both the promise and the current limits of the approach.

What carries the argument

The key machinery is a PCA-derived steering vector: the first principal component computed from per-example differences between hidden activations of biased and neutral sentences at a chosen layer. Injected into forward passes as a scaled addition, it linearly displaces the model's representation space at that layer (L=16), with the steering strength controlled by a coefficient alpha.

What would settle it

Run the same counter-narrative generation with the steering vector removed but with identical prompts and decoding settings; if the official score does not drop significantly, the vector is not the active ingredient. Alternatively, apply the same steering vector to unrelated, non-gender prompts and check whether outputs become generically inclusive but lose relevant content.

Watch

Extended reading notes

Core claim

For counter-narrative generation, the authors compute the first principal component of activation differences between biased and neutral sentence pairs, then add a scaled version of that vector to the hidden states at layer 16 of Gemma-3-4B-it during inference. Combined with strict single-line prompt templates, this steering shifts outputs toward polite, context-aware counter-narratives without modifying model weights, achieving an official average score of 78.12%. The paper also reports that a LoRA fine-tuned model scores 80.00% on the gender-inclusive rewriting subtask.

Load-bearing premise

The approach assumes that gender-inclusivity is encoded as a single linear direction at layer 16; if that attribute is entangled across layers or nonlinear, the injected vector only partially controls the intended behavior.

Editorial extensions

If this is right

Activation steering offers a lightweight, inference-time route to behavioral alignment, avoiding the compute cost of gradient-based fine-tuning.
The method is sensitive to steering strength; the paper recommends a coefficient range (about 0.6-1.2) to balance neutrality against coherence.
Strict prompt templates and decoding settings materially affect judge scores, indicating that prompting and steering interact.
The documented failure modes (semantic drift, residual bias leakage, over-steering, repetition) bound the practical operating envelope of single-layer steering.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If a single linear direction at layer 16 is sufficient for gender-inclusivity, similar attribute-specific vectors could be cheaply derived for other social constraints, though the entanglement observed suggests control may be partial.
An ablation that removes the steering vector while keeping prompts and decoding settings fixed would isolate the vector's contribution from the prompt's; the paper does not report such a control.
The direction is learned from English contrastive pairs, so transfer to other languages may be limited by culture-specific encoding of gender norms.
Combining multi-layer or iterative steering with semantic-preservation constraints could plausibly improve the quality-relevance subscore (64.68%).

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here is my read of 2607.23083. The one thing to know: the paper is an honest, clearly-written shared-task system description, but the causal claim in the abstract — that activation steering produces the 78.12% counter-narrative score — is not supported by the experiments. There is no α=0 no-steering ablation, and given the strict prompt template and low-temperature decoding described in §7, steering may be contributing far less than the framing suggests. That is not a manufactured flaw; the stress-test note lands on reading the paper.

What is new and worth credit: the application of the PCA-difference activation steering protocol (Zou/Turner, properly cited) to gender-inclusive counter-narrative generation, and the manual failure taxonomy on 47 outputs. The five failure modes — semantic drift, residual bias leakage, layer entanglement, over-steering, repetition — ring true and are reported with concrete examples. The authors also do not oversell: they report rank 6 of 7 and the 64.68% QR subscore, and the limitations section admits the single-layer linear-encoding assumption. That is genuine good faith.

The soft spots, in order of severity. The missing baseline is the load-bearing issue. Without a no-steering control under identical prompts and decoding, the paper cannot attribute its official scores to representation engineering. The steering vector, layer, and coefficient are all fitted on the same task distribution, per §4.1 and the notebook, so even the 78.12% is a tuned system result, not a probe of the method. I would also note the manual analysis has no inter-annotator agreement, and the paper does not specify how the 47 cases were sampled. Those are minor for a shared-task report. The linear-encoding worry in §7.1.5 is secondary; the α=0 control would resolve the more basic question.

Bottom line: this is a legitimate system description, not a research breakthrough. If the authors add the no-steering ablation, specify the PCA train/test split, and release the notebook hyperparameters, it becomes a useful record of where activation steering helps and where it collapses. On its own, the scientific claim should be read as a hypothesis, not a result.

My recommendation: engage with it as a borderline-accept system-description paper. Send it to review, and make the α=0 baseline a required revision. The failure taxonomy alone justifies a look.

Referee Report

3 major / 6 minor

Summary. The paper reports the IHLC system submitted to the LT-EDI 2026 shared task on gender-inclusive language generation. For Subtask A (gender-neutral rewriting), the authors fine-tune Gemma-3-4B-it with LoRA and report an official score of 80.00% (rank 3). For Subtask B (counter-narrative generation), the proposed contribution is an inference-time activation-steering method: a PCA direction is computed from hidden-state differences between biased and inclusive sentence pairs, injected into layer 16 of Gemma-3-4B-it during generation, and combined with a strict one-line prompt template and low-temperature decoding. The system achieved an official average of 78.12% (rank 6). The paper also presents a manual analysis of 47 outputs, identifying five failure modes: faithfulness-edit trade-off, layer-specific entanglement, residual bias leakage, over-steering instability, and repetition/degeneration.

Significance. If the causal attribution to activation steering were established, the paper would offer a useful data point: a compute-efficient, weight-free intervention that shifts a modern instruction-tuned model toward polite and inclusive counter-narratives, with code released as a notebook. The honest documentation of failure modes is also a strength. However, the central claim is currently under-supported. The official 78.12% is reported for the full pipeline (prompt template + decoding configuration + steering), and the paper provides no quantitative comparison with the steering hook disabled. Because the steering vector, coefficient, and layer are all tuned on the task before the official evaluation, the result demonstrates a tuned system rather than a validated representation-engineering method. The contribution is promising but needs an ablation to be load-bearing.

major comments (3)

[§5.3, Table 1, §7] The abstract attributes the 78.12% official average to activation steering, but no α=0 (no-steering) baseline is reported under the same prompt template and decoding settings. Section 7 states that strict prompts alone 'improved automated judge scores by reducing undesirable explanatory prefixes that were penalized as irrelevant,' so prompt formatting and decoding may account for a substantial part of the score. To support the central claim, the authors should add an ablation table comparing (i) the full system, (ii) α=0 with the same strict prompt and decoding, and preferably (iii) control interventions such as a random direction or a direction from an unrelated attribute. Without this, the measured score cannot be causally assigned to the steering vector.
[§4.1, §7.1.5] The steering direction is computed via PCA on the task's own biased/neutral contrastive pairs, and the steering coefficient α and layer L are tuned on the task's development data before the official evaluation. This makes the result a tuned system rather than an independent validation of the claim that 'a single PCA-derived direction' suffices for behavioral control. The paper acknowledges in §7.1.5 that single-layer steering assumes a linear attribute encoding, but it does not test this assumption. A concrete test would be to evaluate the same procedure with a direction computed from a held-out or out-of-domain contrastive set, and to compare against a randomly sampled PCA direction of the same norm. This would separate the contribution of the particular steering content from the generic effect of adding a vector to hidden states.
[§7.1] The manual analysis of 47 outputs is presented with percentages (53%, 23%, 11%, 9%, 4%) as if they were reliable error rates, but no inter-annotator agreement or annotation protocol is reported. This is secondary to the main causal claim, but the quantitative presentation overstates the evidentiary weight of what appears to be a single-annotator, small-sample analysis. The authors should either soften the quantitative framing or provide annotation details and agreement scores if available.

minor comments (6)

[§4.1 step 3] The injection is described as applying to 'every token position (or the last token),' which is ambiguous. Section 7.1 says the vector is computed from last-token hidden-state differences. Please specify exactly which token positions receive the steering at inference time.
[§5.2, §7] For reproducibility, the exact prompt templates, the selected layer (L=16), the steering coefficient α actually used for the official run, and the decoding hyperparameters (temperature, repetition penalty) should be stated in the paper rather than only in the linked notebook.
[§7] 'Trial-error' should be 'trial-and-error.'
[Table 1] The column 'N 9 (participants)' / 'N 6 7 (participants)' in Table 1 is visually confusing. Please separate the participation count from the rank column.
[§7.1] Please clarify which set the 47 generated counterfactuals were drawn from (development vs. test) and how they were sampled, since this affects the interpretation of the failure-mode distribution.
[Footnote 1] The GitHub URL is broken across lines ('Gende-r-Inclusive'); check the rendered text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the system is an empirical application of existing activation-steering methods to a shared task, evaluated on an external benchmark.

full rationale

The paper makes no first-principles derivation whose conclusions are equivalent to its inputs. The steering vector is constructed by PCA on contrastive activations (Section 4.1) following the published protocols of Zou et al. and Turner et al., and the resulting outputs are scored by the LT-EDI organizers rather than predicted from the fitting data. Although the steering coefficient, layer, prompts, and decoding settings were tuned on task-related data, this is hyperparameter/configuration tuning of an empirical system, not a definitional reduction: the paper's own failure-mode analysis (e.g., Section 7.1.4) shows generated outputs can diverge substantially from the neutral targets, so the output is not forced by construction. There are no load-bearing self-citations: all methodological citations are to external prior work by other authors. The absence of an alpha=0 ablation weakens causal attribution of the score to steering, but that is an experimental-control issue, not circularity. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method rests on the standard PCA steering pipeline (Zou et al./Turner et al.), with a data-fitted steering direction, a tuned layer and coefficient, and the organizers' hybrid rubric as ground truth. No new entities are invented, but the evaluation's dependence on a single fitted vector and an unstated train/test split of that vector is the main epistemic load.

free parameters (4)

Steering coefficient α = tuned, recommended 0.6–1.2
Scales the injected steering vector; chosen by trial-and-error (§7). Affects over-steering and degeneration.
Steering layer L = 16
Single mid-layer chosen by trial; layer sensitivity acknowledged in §7.1.5.
PCA steering direction v = PC1 of activation differences = computed from task contrastive pairs
Central object of the method; derived via PCA on biased/neutral hidden-state differences (§4.1).
LoRA hyperparameters (rank, alpha, learning rate) = not reported in paper; deferred to notebook
Subtask A's 80% score depends on these settings, but the paper does not state them (§5.2).

assumptions (4)

domain assumption A single linear direction in activation space controls the inclusivity attribute.
The PCA steering protocol assumes a linearly separable encoding of the attribute in last-token hidden states (§4.1, §7.1.5).
domain assumption The organizers' hybrid LLM-as-judge rubric with human oversight is a valid measure of gender inclusivity and counter-narrative quality.
All reported scores are treated as ground truth from the shared task's evaluation framework (§3.2, §5.3).
domain assumption The contrastive pairs used to compute the PCA steering direction are disjoint from the evaluation set.
The paper does not state the train/test split for the activation pairs; if the same examples appear in the reported 47-case test set, scores would be leakage-inflated (§4.1, §7).
standard math PCA over difference vectors is a standard method for finding meaningful activation directions.
The paper relies on standard PCA and linear algebra from Zou et al. (2023); no formal proof is given or needed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LoRA for Gender-Inclusive Rewriting and Activation Steering for Counter-Narrative Generation." pith.science (2026). https://pith.science/paper/QJMV3V7K

@misc{pith2026260723083,
  author       = {Pith},
  title        = {Pith review of: LoRA for Gender-Inclusive Rewriting and Activation Steering for Counter-Narrative Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QJMV3V7K}},
  note         = {Machine review of arXiv:2607.23083}
}

read the original abstract

Gender-inclusive language generation seeks to transform biased text into inclusive alternatives while preserving semantic meaning and contextual coherence. This paper presents the IHLC system for the LT-EDI 2026 Shared Task, addressing both gender-inclusive rewriting and counter-narrative generation. For gender-inclusive rewriting, we employ parameter-efficient Low-Rank Adaptation (LoRA) fine-tuning, achieving an official score of 80.00%. Our primary contribution is a compute-efficient inference-time representation engineering approach for counter-narrative generation. We derive a principal steering direction from contrastive hidden-state activations using principal component analysis (PCA) and inject it into the intermediate representations of Gemma-3-4B-it during inference, enabling behavioral steering toward inclusive responses without modifying model weights. Combined with constrained prompting, this approach produces polite and contextually appropriate counter-narratives, achieving an official score of 78.12%. We further present a manual analysis of steering behavior, identifying key failure modes including semantic drift, residual bias leakage, layer sensitivity, over-steering, and text degeneration. Our findings highlight both the practical potential and current limitations of activation steering as a lightweight alternative to parameter updates for controllable and socially aligned language generation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 5 linked inside Pith

[1]

Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29

2016
[2]

Helena Bonaldi, Serra Sinem Tekiro g lu, and Marco Guerini. 2022. Human-machine collaboration for generating counter-narratives. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4279--4292

2022
[3]

Bharathi Raja Chakravarthi, Shunmuga Priya Muthusamy Chinnan, Paul Buitelaar, Meghann L. Drury-Grogan, Miguel \'A ngel Garc \'i a Cumbreras, Salud Mar \'i a Jim \'e nez Zafra, Thomas Mandl, Sylvia Jaki, Rahul Ponnusamy, Anand Kumar M, Dhanalakshmi V, Bharathi B, Premjith B, Senthil Kumar B, and Sathiyaraj T. 2026. Insights from Multilingual Gender Inclusi...

2026
[4]

Yi-Ling Chung, Serra Sinem Tekiro g lu, and Marco Guerini. 2021. Towards knowledge-grounded counter narrative generation for hate speech. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 899--914

2021
[5]

Gemma Team. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786

arXiv 2025
[6]

Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In International Conference on Learning Representations

2020
[7]

Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36

2024
[8]

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359--17372

2022

Show all 18 references

[9]

Shunmuga Priya Muthusamy Chinnan, Meghann Drury-Grogan, and Bharathi Raja Chakravarthi. 2025. Gender inclusive language generation framework: A reasoning approach with rag and cot. Knowledge-Based Systems, 328:114092

2025
[10]

Jing Qian, Anna Bethke, Yinyin Liu, Elizabeth Belding, and William Yang Wang. 2019. A benchmark dataset for learning to intervene in online hate speech. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 4755--4764

2019
[11]

Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. arXiv preprint arXiv:1909.01326

2019 arXiv
[12]

Nishant Subramani, Nithya Suresh, and Matthew E Peters. 2022. Extracting latent steering vectors from pretrained language models. Findings of the Association for Computational Linguistics: ACL 2022, pages 566--581

2022
[13]

Tony Sun, Andrew Gaut, Tang Shirlyn, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. Mitigating gender bias in natural language processing: Literature review. arXiv preprint arXiv:1906.08976

2019 arXiv
[14]

Serra Sinem Tekiro g lu, Yi-Ling Chung, and Marco Guerini. 2020. Generating counter narratives against online hate speech: Data and strategies. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1177--1190

2020
[15]

Alex Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Maciej MacDiarmid. 2023. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248

2023 arXiv
[16]

Eva Vanmassenhove, Christian Hardmeier, and Andy Way. 2021. Neutral rewriter: A rule-based and neural approach to automatic rewriting into gender neutral alternatives. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8940--8948

2021
[17]

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Hao, Zhanghao Wu, Joseph E Ba, Hao Zhuang, Zi Lin, Zhuohan Li, Eric Xing, and 1 others. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36

2024
[18]

Andy Zou, Long Phan, Sarah Chen, James Campbell, Pengfei Xia, Darren Lin, Minqi Jiang Wang, Danqi Yin, Mantas Woodside, and 1 others. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405

2023 arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29

2016

[2] [2]

Helena Bonaldi, Serra Sinem Tekiro g lu, and Marco Guerini. 2022. Human-machine collaboration for generating counter-narratives. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4279--4292

2022

[3] [3]

Bharathi Raja Chakravarthi, Shunmuga Priya Muthusamy Chinnan, Paul Buitelaar, Meghann L. Drury-Grogan, Miguel \'A ngel Garc \'i a Cumbreras, Salud Mar \'i a Jim \'e nez Zafra, Thomas Mandl, Sylvia Jaki, Rahul Ponnusamy, Anand Kumar M, Dhanalakshmi V, Bharathi B, Premjith B, Senthil Kumar B, and Sathiyaraj T. 2026. Insights from Multilingual Gender Inclusi...

2026

[4] [4]

Yi-Ling Chung, Serra Sinem Tekiro g lu, and Marco Guerini. 2021. Towards knowledge-grounded counter narrative generation for hate speech. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 899--914

2021

[5] [5]

Gemma Team. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786

arXiv 2025

[6] [6]

Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In International Conference on Learning Representations

2020

[7] [7]

Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36

2024

[8] [8]

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359--17372

2022